Convolutional Long Short-Term Memory Network Model for Dynamic Texture Classification: A Case Study

Abstract

The digital world is witnessing the rapid growth of various data such as multimedia content. This enormous volume of information resources is required to be well structured and classified to help machines infer relevant information. Dynamic texture classification is among the essential requirements for multimedia content understanding and plays a vital role in major computer vision applications such as traffic monitoring, face recognition, and surveillance. However, dynamic texture classification brings new challenges to the field of computer vision and has limited literature compared to static texture classification. This paper is another contribution to fill this gap by investigating video classification using textures. In doing so, we have compiled and prepared a dataset of short videos with different textures from various sources, including the DynTex database. These videos are classified according to five categories, namely Clouds/Steam, Fire, Flags, Trees, and Water. Subsequently, the multiclass categorization is performed using the convolutional long short-term memory network (ConvLSTM)-based classifier. The ConvLSTM is a recurrent neural network (RNN) model, just like the LSTM, but the internal matrix multiplications are exchanged with convolution operations. Hence, The ConvLSTM is kind of a combination of Convolution and LSTM. Finally, the experiments have shown that the ConvLSTM succeeded in capturing the spatial features from each texture and making accurate predictions.